def _recreate_variable(self, proto):
name = proto.name if proto.name else None
if name is not None:
dbg_name = name
else:
dbg_name = "<variable loaded from saved model>"
synchronization, aggregation, trainable = (
variables.validate_synchronization_aggregation_trainable(
proto.synchronization, proto.aggregation, proto.trainable,
name=dbg_name))
def uninitialized_variable_creator(next_creator, **kwargs):
"""A variable creator that creates uninitialized variables."""
del next_creator
return resource_variable_ops.UninitializedVariable(**kwargs)
# Create a variable_creator_scope that creates uninitialized variables with
# a lower priority such that a potential distributed variable_creator_scope
# can take precedence.
with ops.get_default_graph()._variable_creator_scope( # pylint: disable=protected-access
uninitialized_variable_creator,
priority=50):
return variables.Variable(
shape=proto.shape,
dtype=proto.dtype,
name=name,
trainable=trainable,
synchronization=synchronization,
aggregation=aggregation), setattr
def _recreate_variable(self, proto):
# TODO(andresp): Can we use the checkpointed value as initializer?
dummy_value = init_ops.Zeros(dtype=proto.dtype)(shape=proto.shape)
synchronization, aggregation, trainable = (
variables.validate_synchronization_aggregation_trainable(
proto.synchronization, proto.aggregation, proto.trainable,
# TODO(allenl): We should save variable names.
name="<variable recreated from SavedModel>"))
return variables.Variable(
dummy_value,
trainable=trainable,
synchronization=synchronization,
aggregation=aggregation), setattr
def _recreate_variable(self, proto):
name = proto.name if proto.name else None
if name is not None:
dbg_name = name
else:
dbg_name = "<variable loaded from saved model>"
synchronization, aggregation, trainable = (
variables.validate_synchronization_aggregation_trainable(
proto.synchronization, proto.aggregation, proto.trainable,
name=dbg_name))
return resource_variable_ops.UninitializedVariable(
shape=proto.shape,
dtype=proto.dtype,
name=name,
trainable=trainable,
synchronization=synchronization,
aggregation=aggregation), setattr
def _recreate_variable(self, proto):
# TODO(andresp): Can we use the checkpointed value as initializer?
dummy_value = init_ops.Zeros(dtype=proto.dtype)(shape=proto.shape)
name = proto.name if proto.name else None
if name is not None:
dbg_name = name
else:
dbg_name = "<variable loaded from saved model>"
synchronization, aggregation, trainable = (
variables.validate_synchronization_aggregation_trainable(
proto.synchronization,
proto.aggregation,
proto.trainable,
name=dbg_name))
return variables.Variable(dummy_value,
name=name,
trainable=trainable,
synchronization=synchronization,
aggregation=aggregation), setattr
def __init__(
self, # pylint: disable=super-init-not-called
initial_value=None,
trainable=None,
caching_device=None,
name=None,
dtype=None,
constraint=None,
add_initializers_to=None,
lifted_initializer_graph=None,
synchronization=None,
aggregation=None,
**unused_kwargs):
"""Creates a variable.
Args:
initial_value: A `Tensor`, or Python object convertible to a `Tensor`,
which is the initial value for the Variable. The initial value must have
a shape specified unless `validate_shape` is set to False. Can also be a
callable with no argument that returns the initial value when called.
(Note that initializer functions from init_ops.py must first be bound
to a shape before being used here.)
trainable: If `True`, GradientTapes automatically watch uses of this
Variable.
caching_device: Optional device string or function describing where the
Variable should be cached for reading. Defaults to the Variable's
device. If not `None`, caches on another device. Typical use is to
cache on the device where the Ops using the Variable reside, to
deduplicate copying through `Switch` and other conditional statements.
name: Optional name for the variable. Defaults to `'Variable'` and gets
uniquified automatically.
dtype: If set, initial_value will be converted to the given type.
If None, either the datatype will be kept (if initial_value is
a Tensor) or float32 will be used (if it is a Python object convertible
to a Tensor).
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
add_initializers_to: if not None and not in legacy graph mode, the
initializer tensor will be added to this map in addition to adding the
assignment to the function.
lifted_initializer_graph: FuncGraph to try to lift initializers to.
synchronization: Indicates when a distributed a variable will be
aggregated. Accepted values are constants defined in the class
`tf.VariableSynchronization`. By default the synchronization is set to
`AUTO` and the current `DistributionStrategy` chooses
when to synchronize. If `synchronization` is set to `ON_READ`,
`trainable` must not be set to `True`.
aggregation: Indicates how a distributed variable will be aggregated.
Accepted values are constants defined in the class
`tf.VariableAggregation`.
Raises:
ValueError: If the initial value is not specified, or does not have a
shape and `validate_shape` is `True`.
RuntimeError: If called outside of a function definition.
"""
if not ops.inside_function():
# If we've been init_scope()d out of the function definition nothing to do
# here; we can't really do the capturing or conditional logic.
resource_variable_ops.ResourceVariable.__init__(
self,
initial_value=initial_value,
trainable=trainable,
caching_device=caching_device,
name=name,
dtype=dtype,
constraint=constraint)
return
with ops.init_scope():
self._in_graph_mode = not context.executing_eagerly()
if initial_value is None:
raise ValueError("initial_value must be specified.")
init_from_fn = callable(initial_value)
if constraint is not None and not callable(constraint):
raise ValueError("The `constraint` argument must be a callable.")
if isinstance(initial_value, trackable.CheckpointInitialValue):
self._maybe_initialize_trackable()
self._update_uid = initial_value.checkpoint_position.restore_uid
initial_value = initial_value.wrapped_value
synchronization, aggregation, trainable = (
variables.validate_synchronization_aggregation_trainable(
synchronization, aggregation, trainable, name))
self._trainable = trainable
self._synchronization = synchronization
self._aggregation = aggregation
self._save_slice_info = None
self._initial_value = None
self._initializer_op = None
self._is_initialized_op = None
self._graph_element = None
self._cached_value = None
# Store the graph key so optimizers know how to only retrieve variables from
# this graph. Guaranteed to be the same as the eager graph_key.
self._graph_key = ops.get_default_graph()._graph_key # pylint: disable=protected-access
with ops.name_scope(name, "Variable",
[] if init_from_fn else [initial_value]) as name:
# pylint: disable=protected-access
with ops.init_scope():
handle_name = ops.name_from_scope_name(name)
unique_id = "%s_%d" % (handle_name, ops.uid())
shared_name = context.shared_name(unique_id)
with ops.name_scope("Initializer"), ops.device(None):
initial_value = ops.convert_to_tensor(
initial_value() if init_from_fn else initial_value,
name="initial_value",
dtype=dtype)
with ops.init_scope():
self._handle = resource_variable_ops.eager_safe_variable_handle(
initial_value=initial_value,
shared_name=shared_name,
name=name,
graph_mode=self._in_graph_mode)
self._shape = initial_value.shape
self._unique_id = unique_id
self._handle_name = handle_name + ":0"
self._dtype = initial_value.dtype.base_dtype
self._constraint = constraint
assert initial_value is not None
if self._in_graph_mode:
with ops.init_scope():
outer_graph = ops.get_default_graph()
func_graph = ops.get_default_graph()
function_placeholders = (func_graph.inputs +
func_graph.internal_captures)
placeholder_ops = set(
[tensor.op for tensor in function_placeholders])
lifted_initializer = lift_to_graph.lift_to_graph(
[initial_value],
outer_graph,
disallowed_placeholders=placeholder_ops)[initial_value]
with ops.init_scope():
self._initial_value = lifted_initializer
with ops.name_scope("IsInitialized"):
self._is_initialized_op = (
resource_variable_ops.var_is_initialized_op(
self._handle))
if initial_value is not None:
with ops.name_scope("Assign") as n, ops.colocate_with(
self._handle):
self._initializer_op = resource_variable_ops.assign_variable_op(
self._handle, lifted_initializer, name=n)
with ops.name_scope("Read"), ops.colocate_with(
self._handle):
# Manually assign reads to the handle's device to avoid log
# messages.
with ops.device(self._handle.device):
value = self._read_variable_op()
self._graph_element = value
ops.add_to_collection(ops.GraphKeys.GLOBAL_VARIABLES, self)
else:
if add_initializers_to is not None:
add_initializers_to[self] = initial_value
def assign_fn():
with ops.name_scope("Assign") as n, ops.colocate_with(
self._handle):
resource_variable_ops.assign_variable_op(self._handle,
initial_value,
name=n)
# Returning values to keep tf.cond happy.
return ops.convert_to_tensor(1)
def not_assign_fn():
return ops.convert_to_tensor(0)
# Note: this cond is always guaranteed to run because we're inside a
# defun which will insert automatic control dependencies.
control_flow_ops.cond(
resource_variable_ops.var_is_initialized_op(self._handle),
not_assign_fn, assign_fn)
# After the handle has been created, set up a way to clean it up when
# executing eagerly. We'll hold the only reference to the deleter, so that
# when this object is garbage collected the deleter will be too. This
# means ResourceVariables can be part of reference cycles without those
# cycles being uncollectable.
if not self._in_graph_mode:
self._handle_deleter = resource_variable_ops.EagerResourceDeleter(
handle=self._handle, handle_device=self._handle.device)
self._cached_shape_as_list = None
def _init_from_args(self,
initial_value=None,
trainable=None,
collections=None,
caching_device=None,
name=None,
dtype=None,
constraint=None,
synchronization=None,
aggregation=None,
distribute_strategy=None,
shape=None):
"""Creates a variable.
Args:
initial_value: A `Tensor`, or Python object convertible to a `Tensor`,
which is the initial value for the Variable. The initial value must have
a shape specified unless `validate_shape` is set to False. Can also be a
callable with no argument that returns the initial value when called.
(Note that initializer functions from init_ops.py must first be bound
to a shape before being used here.)
trainable: If `True`, the default, also adds the variable to the graph
collection `GraphKeys.TRAINABLE_VARIABLES`. This collection is used as
the default list of variables to use by the `Optimizer` classes.
Defaults to `True`, unless `synchronization` is set to `ON_READ`, in
which case it defaults to `False`.
collections: List of graph collections keys. The new variable is added to
these collections. Defaults to `[GraphKeys.GLOBAL_VARIABLES]`.
caching_device: Optional device string or function describing where the
Variable should be cached for reading. Defaults to the Variable's
device. If not `None`, caches on another device. Typical use is to
cache on the device where the Ops using the Variable reside, to
deduplicate copying through `Switch` and other conditional statements.
name: Optional name for the variable. Defaults to `'Variable'` and gets
uniquified automatically.
dtype: If set, initial_value will be converted to the given type.
If None, either the datatype will be kept (if initial_value is
a Tensor) or float32 will be used (if it is a Python object convertible
to a Tensor).
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
synchronization: Indicates when a distributed a variable will be
aggregated. Accepted values are constants defined in the class
`tf.VariableSynchronization`. By default the synchronization is set to
`AUTO` and the current `DistributionStrategy` chooses
when to synchronize.
aggregation: Indicates how a distributed variable will be aggregated.
Accepted values are constants defined in the class
`tf.VariableAggregation`.
distribute_strategy: DistributionStrategy under which this variable
was created.
shape: (optional) The shape of this variable. If None, the shape of
`initial_value` will be used. When setting this argument to
`tf.TensorShape(None)` (representing an unspecified shape), the variable
can be assigned with values of different shapes.
Raises:
ValueError: If the initial value is not specified, or does not have a
shape and `validate_shape` is `True`.
@compatibility(eager)
When Eager Execution is enabled, variables are never added to collections.
It is not implicitly added to the `GLOBAL_VARIABLES` or
`TRAINABLE_VARIABLES` collections, and the `collections` argument is
ignored.
@end_compatibility
"""
synchronization, aggregation, trainable = (
variables.validate_synchronization_aggregation_trainable(
synchronization, aggregation, trainable, name))
if initial_value is None:
raise ValueError("initial_value must be specified.")
init_from_fn = callable(initial_value)
if isinstance(initial_value, ops.Tensor) and hasattr(
initial_value,
"graph") and initial_value.graph.building_function:
raise ValueError(
"Tensor-typed variable initializers must either be "
"wrapped in an init_scope or callable "
"(e.g., `tf.Variable(lambda : "
"tf.truncated_normal([10, 40]))`) when building "
"functions. Please file a feature request if this "
"restriction inconveniences you.")
if collections is None:
collections = [ops.GraphKeys.GLOBAL_VARIABLES]
if not isinstance(collections, (list, tuple, set)):
raise ValueError(
"collections argument to Variable constructor must be a list, tuple, "
"or set. Got %s of type %s" % (collections, type(collections)))
if constraint is not None and not callable(constraint):
raise ValueError("The `constraint` argument must be a callable.")
if isinstance(initial_value, trackable.CheckpointInitialValue):
self._maybe_initialize_trackable()
self._update_uid = initial_value.checkpoint_position.restore_uid
initial_value = initial_value.wrapped_value
if trainable and ops.GraphKeys.TRAINABLE_VARIABLES not in collections:
collections = list(collections) + [
ops.GraphKeys.TRAINABLE_VARIABLES
]
with ops.init_scope():
self._in_graph_mode = not context.executing_eagerly()
with ops.name_scope(
name, "TrainableWrapper",
[] if init_from_fn else [initial_value]) as name:
# pylint: disable=protected-access
handle_name = ops.name_from_scope_name(name)
handle_name = (handle_name or "TrainableWrapperHandle")
if self._in_graph_mode:
shared_name = handle_name
unique_id = shared_name
else:
# When in eager mode use a uid for the shared_name, to prevent
# accidental sharing.
unique_id = "%s_%d" % (handle_name, ops.uid())
shared_name = None #context.shared_name()
# Use attr_scope and device(None) to simulate the behavior of
# colocate_with when the variable we want to colocate with doesn't
# yet exist.
device_context_manager = (ops.device if self._in_graph_mode
else ops.NullContextmanager)
attr = attr_value_pb2.AttrValue(
list=attr_value_pb2.AttrValue.ListValue(
s=[compat.as_bytes("loc:@%s" % handle_name)]))
with ops.get_default_graph()._attr_scope({"_class": attr}):
with ops.name_scope("Initializer"), device_context_manager(
None):
initial_value = ops.convert_to_tensor(
initial_value() if init_from_fn else initial_value,
name="initial_value",
dtype=dtype)
if shape is None:
shape = initial_value.shape
handle = resource_variable_ops.eager_safe_variable_handle(
initial_value=initial_value,
shape=None, # shape,
shared_name=shared_name,
name=name,
graph_mode=self._in_graph_mode)
# pylint: disable=protected-access
if (self._in_graph_mode and initial_value is not None
and initial_value.op._get_control_flow_context()
is not None):
raise ValueError(
"Initializer for variable %s is from inside a control-flow "
"construct, such as a loop or conditional. When creating a "
"variable inside a loop or conditional, use a lambda as the "
"initializer." % name)
# pylint: enable=protected-access
dtype = initial_value.dtype.base_dtype
if self._in_graph_mode:
with ops.name_scope("IsInitialized"):
is_initialized_op = (gen_resource_variable_ops.
var_is_initialized_op(handle))
if initial_value is not None:
# pylint: disable=g-backslash-continuation
with ops.name_scope("Assign") as n, \
ops.colocate_with(None, ignore_existing=True), \
ops.device(handle.device):
# pylint: disable=protected-access
initializer_op = (
gen_resource_variable_ops.assign_variable_op(
handle,
variables.
_try_guard_against_uninitialized_dependencies(
name, initial_value),
name=n))
# pylint: enable=protected-access
# pylint: enable=g-backslash-continuation
with ops.name_scope("Read"):
# Manually assign reads to the handle's device to avoid log
# messages.
with ops.device(handle.device):
with ops.control_dependencies([
gen_resource_variable_ops.
assign_variable_op(
handle,
self.prefetch_values(),
name="AssignBeforeInitRead")
]):
value = gen_resource_variable_ops.read_variable_op(
handle, dtype)
graph_element = value
if caching_device is not None:
# Variables may be created in a tf.device() or ops.colocate_with()
# context. At the same time, users would expect caching device to
# be independent of this context, and/or would not expect the
# current device context to be merged with the caching device
# spec. Therefore we reset the colocation stack before creating
# the cached value. Note that resetting the colocation stack will
# also reset the device stack.
with ops.colocate_with(None, ignore_existing=True):
with ops.device(caching_device):
cached_value = array_ops.identity(value)
else:
cached_value = None
else:
gen_resource_variable_ops.assign_variable_op(
handle, initial_value)
is_initialized_op = None
initializer_op = None
graph_element = None
if caching_device:
with ops.device(caching_device):
with ops.control_dependencies([
gen_resource_variable_ops.
assign_variable_op(
handle,
self.prefetch_values(),
name="AssignBeforeInitRead")
]):
cached_value = gen_resource_variable_ops.read_variable_op(
handle, dtype)
else:
cached_value = None
if not context.executing_eagerly():
# Eager variables are only added to collections if they are part of an
# eager variable store (otherwise in an interactive session they would
# hog memory and cause OOM). This is done in ops/variable_scope.py.
ops.add_to_collections(collections, self)
elif ops.GraphKeys.GLOBAL_STEP in collections:
ops.add_to_collections(ops.GraphKeys.GLOBAL_STEP, self)
initial_value = initial_value if self._in_graph_mode else None
super(resource_variable_ops.ResourceVariable,
self).__init__(trainable=trainable,
shape=shape,
dtype=dtype,
handle=handle,
synchronization=synchronization,
constraint=constraint,
aggregation=aggregation,
distribute_strategy=distribute_strategy,
name=name,
unique_id=unique_id,
handle_name=handle_name,
graph_element=graph_element,
initial_value=initial_value,
initializer_op=initializer_op,
is_initialized_op=is_initialized_op,
cached_value=cached_value)
def __init__(self, # pylint: disable=super-init-not-called
initial_value=None,
trainable=None,
caching_device=None,
name=None,
dtype=None,
constraint=None,
add_initializers_to=None,
lifted_initializer_graph=None,
synchronization=None,
aggregation=None,
**unused_kwargs):
"""Creates a variable.
Args:
initial_value: A `Tensor`, or Python object convertible to a `Tensor`,
which is the initial value for the Variable. The initial value must have
a shape specified unless `validate_shape` is set to False. Can also be a
callable with no argument that returns the initial value when called.
(Note that initializer functions from init_ops.py must first be bound
to a shape before being used here.)
trainable: If `True`, GradientTapes automatically watch uses of this
Variable.
caching_device: Optional device string or function describing where the
Variable should be cached for reading. Defaults to the Variable's
device. If not `None`, caches on another device. Typical use is to
cache on the device where the Ops using the Variable reside, to
deduplicate copying through `Switch` and other conditional statements.
name: Optional name for the variable. Defaults to `'Variable'` and gets
uniquified automatically.
dtype: If set, initial_value will be converted to the given type.
If None, either the datatype will be kept (if initial_value is
a Tensor) or float32 will be used (if it is a Python object convertible
to a Tensor).
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value
(which must have the same shape). Constraints are not safe to
use when doing asynchronous distributed training.
add_initializers_to: if not None and not in legacy graph mode, the
initializer tensor will be added to this map in addition to adding the
assignment to the function.
lifted_initializer_graph: FuncGraph to try to lift initializers to.
synchronization: Indicates when a distributed a variable will be
aggregated. Accepted values are constants defined in the class
`tf.VariableSynchronization`. By default the synchronization is set to
`AUTO` and the current `DistributionStrategy` chooses
when to synchronize. If `synchronization` is set to `ON_READ`,
`trainable` must not be set to `True`.
aggregation: Indicates how a distributed variable will be aggregated.
Accepted values are constants defined in the class
`tf.VariableAggregation`.
Raises:
ValueError: If the initial value is not specified, or does not have a
shape and `validate_shape` is `True`.
RuntimeError: If called outside of a function definition.
"""
if not ops.inside_function():
# If we've been init_scope()d out of the function definition nothing to do
# here; we can't really do the capturing or conditional logic.
resource_variable_ops.ResourceVariable.__init__(
self, initial_value=initial_value, trainable=trainable,
caching_device=caching_device, name=name, dtype=dtype,
constraint=constraint)
return
with ops.init_scope():
self._in_graph_mode = not context.executing_eagerly()
if initial_value is None:
raise ValueError("initial_value must be specified.")
init_from_fn = callable(initial_value)
if constraint is not None and not callable(constraint):
raise ValueError("The `constraint` argument must be a callable.")
if isinstance(initial_value, trackable.CheckpointInitialValue):
self._maybe_initialize_trackable()
self._update_uid = initial_value.checkpoint_position.restore_uid
initial_value = initial_value.wrapped_value
synchronization, aggregation, trainable = (
variables.validate_synchronization_aggregation_trainable(
synchronization, aggregation, trainable, name))
self._trainable = trainable
self._synchronization = synchronization
self._aggregation = aggregation
self._save_slice_info = None
self._initial_value = None
self._initializer_op = None
self._is_initialized_op = None
self._graph_element = None
self._cached_value = None
# Store the graph key so optimizers know how to only retrieve variables from
# this graph. Guaranteed to be the same as the eager graph_key.
self._graph_key = ops.get_default_graph()._graph_key # pylint: disable=protected-access
with ops.name_scope(name, "Variable", []
if init_from_fn else [initial_value]) as name:
# pylint: disable=protected-access
with ops.init_scope():
handle_name = ops.name_from_scope_name(name)
unique_id = "%s_%d" % (handle_name, ops.uid())
shared_name = context.shared_name(unique_id)
with ops.name_scope("Initializer"), ops.device(None):
initial_value = ops.convert_to_tensor(
initial_value() if init_from_fn else initial_value,
name="initial_value", dtype=dtype)
with ops.init_scope():
self._handle = resource_variable_ops.eager_safe_variable_handle(
initial_value=initial_value,
shared_name=shared_name,
name=name,
graph_mode=self._in_graph_mode)
self._shape = initial_value.shape
self._unique_id = unique_id
self._handle_name = handle_name + ":0"
self._dtype = initial_value.dtype.base_dtype
self._constraint = constraint
assert initial_value is not None
if self._in_graph_mode:
with ops.init_scope():
outer_graph = ops.get_default_graph()
func_graph = ops.get_default_graph()
function_placeholders = (
func_graph.inputs + func_graph.internal_captures)
placeholder_ops = set(
[tensor.op for tensor in function_placeholders])
lifted_initializer = lift_to_graph.lift_to_graph(
[initial_value], outer_graph,
disallowed_placeholders=placeholder_ops)[initial_value]
with ops.init_scope():
self._initial_value = lifted_initializer
with ops.name_scope("IsInitialized"):
self._is_initialized_op = (
resource_variable_ops.var_is_initialized_op(self._handle))
if initial_value is not None:
with ops.name_scope("Assign") as n, ops.colocate_with(self._handle):
self._initializer_op = resource_variable_ops.assign_variable_op(
self._handle, lifted_initializer, name=n)
with ops.name_scope("Read"), ops.colocate_with(self._handle):
# Manually assign reads to the handle's device to avoid log
# messages.
with ops.device(self._handle.device):
value = self._read_variable_op()
self._graph_element = value
ops.add_to_collection(ops.GraphKeys.GLOBAL_VARIABLES, self)
else:
if add_initializers_to is not None:
add_initializers_to[self] = initial_value
def assign_fn():
with ops.name_scope("Assign") as n, ops.colocate_with(self._handle):
resource_variable_ops.assign_variable_op(
self._handle,
initial_value,
name=n)
# Returning values to keep tf.cond happy.
return ops.convert_to_tensor(1)
def not_assign_fn():
return ops.convert_to_tensor(0)
# Note: this cond is always guaranteed to run because we're inside a
# defun which will insert automatic control dependencies.
control_flow_ops.cond(
resource_variable_ops.var_is_initialized_op(self._handle),
not_assign_fn, assign_fn)
# After the handle has been created, set up a way to clean it up when
# executing eagerly. We'll hold the only reference to the deleter, so that
# when this object is garbage collected the deleter will be too. This
# means ResourceVariables can be part of reference cycles without those
# cycles being uncollectable.
if not self._in_graph_mode:
self._handle_deleter = resource_variable_ops.EagerResourceDeleter(
handle=self._handle, handle_device=self._handle.device)
self._cached_shape_as_list = None
def get_variable(
self,
name,
shape=None, # embedding_dim
dtype=dtypes.float32,
ktype=dtypes.int64,
initializer=None,
regularizer=None,
reuse=None,
trainable=None,
collections=None,
caching_device=None,
partitioner=None,
validate_shape=True,
constraint=None,
synchronization=VariableSynchronization.AUTO,
aggregation=VariableAggregation.NONE):
"""Gets an existing variable with these parameters or create a new one.
If a variable with the given name is already stored, we return the stored
variable. Otherwise, we create a new one.
Set `reuse` to `True` when you only want to reuse existing Variables.
Set `reuse` to `False` when you only want to create new Variables.
Set `reuse` to None (the default) or tf.compat.v1.AUTO_REUSE when you want
variables to be created if they don't exist or returned if they do.
If initializer is `None` (the default), the default initializer passed in
the constructor is used. If that one is `None` too, we use a new
`glorot_uniform_initializer`. If initializer is a Tensor, we use
it as a value and derive the shape from the initializer.
If a partitioner is provided, a `PartitionedVariable` is returned.
Accessing this object as a `Tensor` returns the shards concatenated along
the partition axis.
Some useful partitioners are available. See, e.g.,
`variable_axis_size_partitioner` and `min_max_variable_partitioner`.
Args:
name: The name of the new or existing variable.
shape: Shape of the new or existing variable.
dtype: Type of the new or existing variable (defaults to `DT_FLOAT`).
initializer: Initializer for the variable.
regularizer: A (Tensor -> Tensor or None) function; the result of applying
it on a newly created variable will be added to the collection
GraphKeys.REGULARIZATION_LOSSES and can be used for regularization.
reuse: a Boolean, None, or tf.AUTO_REUSE. Controls reuse or creation of
variables. When eager execution is enabled this argument is always
forced to be False.
trainable: If `True` also add the variable to the graph collection
`GraphKeys.TRAINABLE_VARIABLES` (see `tf.Variable`). `trainable`
defaults to `True`, unless `synchronization` is set to `ON_READ`, in
which case it defaults to `False`.
collections: List of graph collections keys to add the `Variable` to.
Defaults to `[GraphKeys.GLOBAL_VARIABLES]` (see `tf.Variable`).
caching_device: Optional device string or function describing where the
Variable should be cached for reading. Defaults to the Variable's
device. If not `None`, caches on another device. Typical use is to
cache on the device where the Ops using the `Variable` reside, to
deduplicate copying through `Switch` and other conditional statements.
partitioner: Optional callable that accepts a fully defined `TensorShape`
and dtype of the `Variable` to be created, and returns a list of
partitions for each axis (currently only one axis can be partitioned).
validate_shape: If False, allows the variable to be initialized with a
value of unknown shape. If True, the default, the shape of initial_value
must be known.
use_resource: If False, creates a regular Variable. If True, creates
instead an experimental ResourceVariable which has well-defined
semantics. Defaults to False (will later change to True). When eager
execution is enabled this argument is always forced to be true.
custom_getter: Callable that takes as a first argument the true getter,
and allows overwriting the internal get_variable method. The signature
of `custom_getter` should match that of this method,
but the most future-proof version will allow for changes: `def
custom_getter(getter, *args, **kwargs)`. Direct access to
all `get_variable` parameters is also allowed: `def
custom_getter(getter, name, *args, **kwargs)`. A simple identity
custom getter that simply creates variables with modified names is:
```python
def custom_getter(getter, name, *args, **kwargs): return getter(name +
'_suffix', *args, **kwargs) ```
constraint: An optional projection function to be applied to the variable
after being updated by an `Optimizer` (e.g. used to implement norm
constraints or value constraints for layer weights). The function must
take as input the unprojected Tensor representing the value of the
variable and return the Tensor for the projected value (which must have
the same shape). Constraints are not safe to use when doing asynchronous
distributed training.
synchronization: Indicates when a distributed a variable will be
aggregated. Accepted values are constants defined in the class
`tf.VariableSynchronization`. By default the synchronization is set to
`AUTO` and the current `DistributionStrategy` chooses when to
synchronize.
aggregation: Indicates how a distributed variable will be aggregated.
Accepted values are constants defined in the class
`tf.VariableAggregation`.
Returns:
The created or existing `Variable` (or `PartitionedVariable`, if a
partitioner was used).
Raises:
ValueError: when creating a new variable and shape is not declared,
when reusing a variable and specifying a conflicting shape,
or when violating reuse during variable creation.
RuntimeError: when eager execution is enabled and not called from an
EagerVariableStore.
"""
# Note that it's fine to reuse eager variables whose initialization was
# lifted from a function-building graph into the eager context (that's why
# the following clause is not wrapped in an `init_scope`); lifted variables
# are tracked by the graph's `VariableStore`.
if context.executing_eagerly():
if not self._var_store._store_eager_variables and reuse:
raise RuntimeError(
"When eager execution is enabled variable reuse is only supported"
" when an EagerVariableStore is active. See the documentation on"
" EagerVariableStore for example usage.")
if self._var_store._store_eager_variables:
reuse = AUTO_REUSE
# If a *_ref type is passed in an error would be triggered further down the
# stack. We prevent this using base_dtype to get a non-ref version of the
# type, before doing anything else. When _ref types are removed in favor of
# resources, this line can be removed.
try:
dtype = dtype.base_dtype
ktype = ktype.base_dtype
except AttributeError:
# .base_dtype not existing means that we will try and use the raw dtype
# which was passed in - this might be a NumPy type which is valid.
pass
# This is the main logic of get_variable. However, custom_getter
# may override this logic. So we save it as a callable and pass
# it to custom_getter.
# Note: the parameters of _true_getter, and their documentation, match
# *exactly* item-for-item with the docstring of this method.
def _true_getter( # pylint: disable=missing-docstring
var_store,
name,
shape=None,
dtype=dtypes.float32,
ktype=dtypes.int64,
initializer=None,
regularizer=None,
reuse=None,
trainable=None,
collections=None,
caching_device=None,
partitioner=None,
validate_shape=True,
constraint=None,
synchronization=VariableSynchronization.AUTO,
aggregation=VariableAggregation.NONE):
is_scalar = (shape is not None
and isinstance(shape, collections_abc.Sequence)
and not shape)
# Partitioned variable case
if partitioner is not None and not is_scalar:
if not callable(partitioner):
raise ValueError(
"Partitioner must be callable, but received: %s" %
partitioner)
with ops.name_scope(None):
return self._get_partitioned_variable(
var_store=var_store,
name=name,
shape=shape,
dtype=dtype,
ktype=ktype,
initializer=initializer,
regularizer=regularizer,
reuse=reuse,
trainable=trainable,
collections=collections,
caching_device=caching_device,
partitioner=partitioner,
validate_shape=validate_shape,
constraint=constraint,
synchronization=synchronization,
aggregation=aggregation)
# Special case for partitioned variable to allow reuse without having to
# specify partitioner.
if (reuse is True and partitioner is None
and name in self._var_store._partitioned_vars):
return self._get_partitioned_variable(
var_store=var_store,
name=name,
shape=shape,
dtype=dtype,
ktype=dtypes.int64,
initializer=initializer,
regularizer=regularizer,
reuse=reuse,
trainable=trainable,
collections=collections,
caching_device=caching_device,
partitioner=None,
validate_shape=validate_shape,
constraint=constraint,
synchronization=synchronization,
aggregation=aggregation)
# Single variable case
if "%s/part_0" % name in var_store._vars:
raise ValueError(
"No partitioner was provided, but a partitioned version of the "
"variable was found: %s/part_0. Perhaps a variable of the same "
"name was already created with partitioning?" % name)
return self._get_single_variable(var_store=var_store,
name=name,
shape=shape,
dtype=dtype,
ktype=ktype,
initializer=initializer,
regularizer=regularizer,
reuse=reuse,
trainable=trainable,
collections=collections,
caching_device=caching_device,
validate_shape=validate_shape,
constraint=constraint,
synchronization=synchronization,
aggregation=aggregation)
synchronization, aggregation, trainable = (
variables.validate_synchronization_aggregation_trainable(
synchronization, aggregation, trainable, name))
return _true_getter(self._var_store,
name,
shape=shape,
dtype=dtype,
ktype=ktype,
initializer=initializer,
regularizer=regularizer,
reuse=reuse,
trainable=trainable,
collections=collections,
caching_device=caching_device,
partitioner=partitioner,
validate_shape=validate_shape,
constraint=constraint,
synchronization=synchronization,
aggregation=aggregation)
